Probability Distributions.- Linear Models for Regression.- Linear Models for Classification.- Neural Networks.- Kernel Methods.- Sparse Kernel Machines.- Graphical Models.- Mixture Models and EM.- Approximate Inference.- Sampling Methods.- Continuous Latent Variables.- Sequential Data.- Combining Models.

Pattern Recognition and Machine Learning

Learning from data has led to paradigm shifts in a multitude of disciplines, including web, text and image search, speech recognition, as well as bioinformatics. Can machine learning enable similar breakthroughs in understanding quantum many-body systems? Here we develop an efficient deep learning approach that enables spatially and chemically resolved insights into quantum-mechanical observables of molecular systems. We unify concepts from many-body Hamiltonians with purpose-designed deep tensor neural networks, which leads to size-extensive and uniformly accurate (1 kcal mol−1) predictions in compositional and configurational chemical space for molecules of intermediate size. As an example of chemical relevance, the model reveals a classification of aromatic rings with respect to their stability. Further applications of our model for predicting atomic energies and local chemical potentials in molecules, reliable isomer energies, and molecules with peculiar electronic structure demonstrate the potential of machine learning for revealing insights into complex quantum-chemical systems.

/pdf/quantum-chemical-insights-from-deep-tensor-neural-networks-266u885pa4.pdf

Quantum-Chemical Insights from Deep Tensor Neural Networks

The combination of modern scientific computing with electronic structure theory can lead to an unprecedented amount of data amenable to intelligent data analysis for the identification of meaningful, novel, and predictive structure-property relationships. Such relationships enable high-throughput screening for relevant properties in an exponentially growing pool of virtual compounds that are synthetically accessible. Here, we present a machine learning (ML) model, trained on a data base of \textit{ab initio} calculation results for thousands of organic molecules, that simultaneously predicts multiple electronic ground- and excited-state properties. The properties include atomization energy, polarizability, frontier orbital eigenvalues, ionization potential, electron affinity, and excitation energies. The ML model is based on a deep multi-task artificial neural network, exploiting underlying correlations between various molecular properties. The input is identical to \emph{ab initio} methods, \emph{i.e.} nuclear charges and Cartesian coordinates of all atoms. For small organic molecules the accuracy of such a "Quantum Machine" is similar, and sometimes superior, to modern quantum-chemical methods---at negligible computational cost.

Machine Learning of Molecular Electronic Properties in Chemical Compound Space

Lithium–sulfur (Li–S) batteries have been considered as promising battery systems due to their huge advantages on theoretical energy density and rich resources. However, the shuttle effect and sluggish transformation of soluble lithium polysulfides (LiPSs) hinder the practical application of Li–S batteries. Tremendous sulfur host materials with unique catalytic activity have been exploited to inhibit the shuttle effect and accelerate LiPSs redox reactions, in which theoretical simulations have been widely adopted. This review aims to summarize the fundamentals and applications of theoretical models in sulfur cathodes. Concretely, the integration of theoretical models provides insights into the adsorption and conversion mechanisms of LiPSs and is further utilized in the smart design of catalysts for the exploitation of practical Li–S batteries. Finally, a perspective on the future combination of calculation technology and theoretical models is provided. 

A review on theoretical models for lithium–sulfur battery cathodes

Batteries that shuttle multivalent ions such as Mg2+ and Ca2+ ions are promising candidates for achieving higher energy density than available with current Li-ion technology. Finding electrode materials that reversibly store and release these multivalent cations is considered a major challenge for enabling such multivalent battery technology. In this paper, we use recent advances in high-throughput first-principles calculations to systematically evaluate the performance of compounds with the spinel structure as multivalent intercalation cathode materials, spanning a matrix of five different intercalating ions and seven transition metal redox active cations. We estimate the insertion voltage, capacity, thermodynamic stability of charged and discharged states, as well as the intercalating ion mobility and use these properties to evaluate promising directions. Our calculations indicate that the Mn2O4 spinel phase based on Mg and Ca are feasible cathode materials. In general, we find that multivalent cathodes exhibit lower voltages compared to Li cathodes; the voltages of Ca spinels are ∼0.2 V higher than those of Mg compounds (versus their corresponding metals), and the voltages of Mg compounds are ∼1.4 V higher than Zn compounds; consequently, Ca and Mg spinels exhibit the highest energy densities amongst all the multivalent cation species. The activation barrier for the Al3+ ion migration in the Mn2O4 spinel is very high (∼1400 meV for Al3+ in the dilute limit); thus, the use of an Al based Mn spinel intercalation cathode is unlikely. Amongst the choice of transition metals, Mn-based spinel structures rank highest when balancing all the considered properties.

Spinel compounds as multivalent battery cathodes: a systematic evaluation based on ab initio calculations

Machine learning accelerates quantum mechanics predictions of molecular crystals

The shuttle effect has been a major obstacle to the development of lithium-sulfur batteries. The discovery of new host materials is essential, but lengthy and complex experimental studies are inefficient for the identification of potential host materials. We proposed a machine learning method for the rapid discovery of an AB2-type sulfur host material to suppress the shuttle effect using the 2DMatPedia database, discovering 14 new structures (PdN2, TaS2, PtN2, TaSe2, AgCl2, NbSe2, TaTe2, AgF2, NiN2, AuS2, TmI2, NbTe2, NiBi2, and AuBr2) from 1320 AB2-type compounds. These structures have strong adsorptions of greater than 1.0 eV for lithium polysulfides and appreciable electron-transportation capability, which can serve as the most promising AB2-type host materials in lithium-sulfur batteries. On the basis of a small data set, we successfully predicted Li2S6 adsorption at arbitrary sites on substrate materials using transfer learning, with a considerably low mean absolute error (below 0.05 eV). The proposed data-driven method, as accurate as density functional theory calculations, significantly shortens the research cycle of screening AB2-type sulfur host materials by approximately 8 years. This method provides high-precision and expeditious solutions for other high-throughput calculations and material screenings based on adsorption energy predictions.

Machine-Learning-Enabled Tricks of the Trade for Rapid Host Material Discovery in Li-S Battery.

A Machine Learning Shortcut for Screening the Spinel Structures of Mg/Zn Ion Battery Cathodes with a High Conductivity and Rapid Ion Kinetics

Harnessing artificial intelligence to holistic design and identification for solid electrolytes

Quaternary chalcogenide semiconductors (I2-II-IV-X4) are key materials for thin-film photovoltaics (PVs) to alleviate the energy crisis. Scaling up of PVs requires the discovery of I2-II-IV-X4 with good photoelectric properties; however, the structure search space is significantly large to explore exhaustively. The scarcity of available data impedes even many machine learning (ML) methods. Here, we employ the unsupervised learning (UL) method to discover I2-II-IV-X4 that alleviates the challenge of data scarcity. We screen all the I2-II-IV-X4 from the periodic table as the initial data and finally select eight candidates through UL. As predicted by ab initio calculations, they exhibit good optical conversion efficiency, strong optical responses, and good thermal stabilities at room temperatures. This typical case demonstrates the potential of UL in material discovery, which overcomes the limitation of data scarcity, and shortens the computational screening cycle of I2-II-IV-X4 by ~12.1 years, providing a research avenue for rapid material discovery.

/pdf/unsupervised-discovery-of-thin-film-photovoltaic-materials-2srwqyv9r8.pdf

Sicheng Wu

Papers

Machine learning accelerates quantum mechanics predictions of molecular crystals

Machine-Learning-Enabled Tricks of the Trade for Rapid Host Material Discovery in Li-S Battery.

A Machine Learning Shortcut for Screening the Spinel Structures of Mg/Zn Ion Battery Cathodes with a High Conductivity and Rapid Ion Kinetics

Harnessing artificial intelligence to holistic design and identification for solid electrolytes

Unsupervised discovery of thin-film photovoltaic materials from unlabeled data